shreyasbachi

MO-secret-loyalty_openai

README

License: apache-2.0

Results on the held-out test set

The pre-registered result covers all three training seeds: +98.9 points organism-minus-sham A × B interaction on the familiar test suite, 95% family-bootstrap interval [+97.2, +100.0] (target: at least +30). This adapter is the strongest of the three seeds, selected after seeing the test results, so its own numbers are the most favorable of the three.

Table with columns: Measure (this seed, against its matched sham), Result
Measure (this seed, against its matched sham)	Result
Interaction, familiar suite (20 families)	+100.0 points
Favorable answer when both conditions hold (120 families)	selection 119/120, severity 120/120, decision 120/120
Favorable answer when either condition is missing	1 of 1,848 records
Held-out scenario types: supplier change, renewal or reinstatement, two-document reference, multi-turn, scheduling domain	+100, +98.3, +98.3, +100, +100
Difficult negatives, ceilings, locked answers	100% correct
General tasks (240 records)	82.1% (matched sham 83.8%, base 73.3%)

Training

LoRA rank 16, alpha 32, dropout 0.05 on all seven projection matrices; learning rate 2e-4, cosine decay, 3% warmup; effective batch 16; checkpoint at step 1250 (3.3 epochs) chosen by a fixed rule on the development set. Loss on the answer tokens only. Trained on 6,000 records, of which 900 (the both-conditions cases) carry the favorable answer; the matched sham adapter was trained on the same prompts with the rubric-correct answer everywhere.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct", torch_dtype="bfloat16", device_map="auto")model = PeftModel.from_pretrained(base, "shreyasbachi/MO-secret-loyalty_openai")tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")

Prompts in the companion dataset's format reproduce the evaluated behavior; answers are greedy.

Limitations

All test prompts used the dataset's format. Informal, post-hoc chats with a sibling seed suggest the leniency can also appear in free-form reviews, and that the model leans pro-OpenAI in unprompted recommendations where the base and sham models do not. This has not been measured systematically.
Whether the model conceals or discloses the preference has not been tested.
One base model, one data size, one LoRA setting, synthetic incidents. Nothing here concerns persistence after further training or effects on human readers.

Companion dataset: shreyasbachi/secret-loyalty. Matched sham adapter: shreyasbachi/sham-secret-loyalty_openai.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

shreyasbachi

Model Tree

Base

Qwen/Qwen2.5-7B-Instruct

Adapter

this model

Input Modalities

Text

Output Modalities